Sampling for computational efficiency when conducting analyses in big data

A challenge to research in big data is the inherent computational intensity of analyses, particularly when using rigorous methods to address biases. We demonstrate the use of sampling methods in big data to estimate parameters using fewer resources. Our motivating question was whether lung cancer incidence differs by baseline HIV status, using a cohort of nearly 30 million Medicaid beneficiaries. We targeted three parameters (with listed estimator): incidence rate ratio (IRR, Poisson model), hazard ratio (HR, Cox model), and risk ratio (RR, Kaplan-Meier). We controlled for confounders using inverse probability weighting. We ran analyses using the full sample and several sampling schemes: divide-and-recombine (10, 20, 50 samples), sub-cohort, and case-cohort. We compared point estimates, standard errors, computation time, and memory used. We observed 1113 incident lung cancer diagnoses among 180,980 beneficiaries with HIV and 33,106 diagnoses among 29,179,940 beneficiaries without HIV. Findings were similar across target parameters. The sub-cohort and case-cohort approaches had estimates closer to the full sample and were faster and less memory-intensive than divide-and-recombine, especially when estimating the RR. Including non-sampled cases in the case-cohort resulted in increases in computation time and memory relative to the sub-cohort approach.